Problem

  • Decoupling disk from data storage

 For large-scale systems, storage is far away from compute

 Relatively lower bandwidth between compute and storage results in higher latency

 This is also dependant on applications

  • Challenging research problem:

 How to evaluate a large-scale system is challengeable

Users

  • Data centre providers

 Need to improve EDP and reduce costs

  • Researchers

 Need a solid methodology to evaluate the performance of large-scale systems

Performance

  • In some cases, whether the performance is improved has a direct impact on users
  • What bounds the performance of large-scale systems?

  • Cost breakdown

 Cores ( compute )

 Network bandwidth

 Memory

 Storage

 Infrastructure ( cooling, power delivery, building )

 Energy cost

 Maintenance ( repairing )

 Software ( a huge cost )

  • Systems may go about 2 to 5 years

  • Maintenance

 5 years (4 * 104 hrs), $0.1 per KW/hr → $ 4000 /KW

  • Recurring Cost

 100 cores / KW

 DRAM → 1 TB / KW ( memory chip < 1 W, about 1 GB per chip )

 Disk → 100 TB / KW ( 1~10 Watts / disk, about 1 TB per disk )

 NVM → 10 TB / KW

 DRAMNVMDisk
Static power OnRefresh + interfacesInterfacesSpinning + Interfaces
Static power OffRefresh\\
Access RActivate array, senseActivate, senseMove ahead, sense
Access WActivateState switchNoise
Move dataHigher bandwidthLonger distances
commCost more

  • Network (not cheap)

 20% overhead

 10% building

 10% cooling

  • Non-recurring Cost

 Core → $ 10 ~ 100 each

 DRAM → $10 / GB

 Disk → $ 0.1 / GB

 Network → $ 1 /share

  • With $ 1000 to spend,

 For cores, corresponding to 10 cores, it costs 100 W

 For DRAMs, corresponding to 100 GB, it costs 100 W

 For Disks, corresponding to 10 TB, it costs 100 W

 For NVMs, corresponding to 1 TB, it costs 100 W

  • Much cost with the existence of disks and DRAMs

 Cost savings when decoupling disks and DRAMs from storage

  • Also improve the performance

 Data stored in NVMs stacked on chips close to compute, decreasing the latency

  • A lot of money put to memory to allow high parallelism

 There should be a balance between cost and efficiency